Erebus-RP-Mini-4B-Instruct-2608-v0.1: Experimental roleplay finetune of Qwen 3.5 4B.
5193119660419783157
Straight to quick overview:
This model was aggressively trained for engaging roleplay, with a methodology slightly unique from my other finetunes, which, not without a slight stability hit, resulted in:
Better creative writing
More fun and consistent roleplay
Improved punctuation
Overhauled writing style
Quants(This time in the repo, with safetensors in the separate folder):
BF16: Overkill
Q8_0: Lossless
Q6_K: Near lossless
Q5_K_M: Very high quality
Q4_K_M: High quality and fastest
Note:
Since this is my experimental attempt, please report any issues you might encounter with it and general feedback.
I also might stop actively fine tuning for a few days or even weeks as I am transistioning back to information gathering in order to improve my SFT dataset. This, however, doesn't mean that I will not accept any feedback.
UPD: caliperbench.com results included: ranks 6th for ERP in 9B and under category
